def check_gradients(x, model=None, delta=1e-10):
"""Compute gradients using symmetric differences.
x: The expression to compute gradients of.
model: Optional list of parameters to compute gradients with respect to.
This is extremely slow and is used only for debugging."""
if model is None:
model = [subx for subx in expr.topological(x) if isinstance(subx, expr.parameter)]
gradients = {}
for param in model:
g = numpy.empty_like(param.value)
it = numpy.nditer([param.value, g], [], [['readwrite'], ['writeonly']])
for pi, gi in it:
save = pi
pi -= delta/2
val_minus = compute_values(x)[x]
pi += delta
val_plus = compute_values(x)[x]
pi[...] = save
gi[...] = (val_plus-val_minus)/delta
gradients[param] = g
return gradients
def compute_values(x, initvalues={}):
"""Evaluate an expression and all its subexpressions.
x: The expression to evaluate.
initvalues: Optional dictionary from subexpressions to
precomputed values; can be used to continue a
computation when the expression grows.
"""
values = collections.OrderedDict()
for subx in expr.topological(x):
if subx in initvalues:
values[subx] = initvalues[subx]
else:
try:
subx.forward(values)
except:
if logging.debug:
sys.stderr.write("Expression traceback (most recent call last):\n" + "".join(logging.format_list(subx.stack)))
raise
if logging.trace:
sys.stdout.write("<%s> = %s = %s" % (subx.serial, subx, format_value(values[subx])))
return values